https://retellai.com/?utm_source=discord&utm_medium=seo logo
Powered by
i Retell team, I’m opening a ticket because I’m s...
# ❓・ai-evy-support
f
i Retell team, @AI Evy I’m opening a ticket because I’m seeing a mismatch between the MCP tool-call result and the assistant’s final reply. Context When I put my secretary agent in Chat / Chatbot mode, it performs the MCP calls correctly. The MCP call returns the expected payload/result, and I can see the tool-call completing successfully. However, the assistant message that follows is hallucinated and does not use the MCP response (it replies with something unrelated, as if it ignored the tool output). Questions Why can the MCP call succeed but the assistant still produces an unrelated response afterward? What are the most likely causes (e.g., tool result not being injected into the model context, wrong message roles, truncation, streaming timing/race conditions, memory/summary interfering, etc.)? What’s the recommended fix to ensure the assistant is forced to ground the final response on the MCP result? What I’ve verified MCP call executes and returns successfully (no error). The problem is specifically the post-tool natural-language response not matching the returned data.
a
Thanks for tagging me @Fernandiyo. I'll look through relevant sources and get back to you shortly.
### Answer Based on the available sources, I don't have direct documentation explaining why an MCP call would succeed but the assistant ignores the result in Chat/Chatbot mode specifically. However, here are some relevant points and recommendations: Prompt Guidance The documentation recommends explicitly instructing the agent when to invoke MCP tools and how to use results. You should include in your prompt when and how to use the tool response: [(1) ]()
Copy code
When user states his name and phone number, please call tool `verify_user` function.
Response Variables You can extract values from the MCP tool response and save them as dynamic variables for use later in the conversation. For example, you can extract a user's name from the response and reference it later using 
{{user_name}}
. Potential Causes to Investigate - Ensure your prompt explicitly instructs the agent to use the MCP response data in its reply - Check if response variables are properly configured to extract and store the MCP result - Verify the prompt includes clear instructions like "After calling [tool_name], use the returned data to answer the user" Known Issues There have been reported issues where agents stop responding correctly after tool calls, though these were typically related to error states like "An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'." [(2) ]() [(3) ]() Your case seems different since the call succeeds without error. Recommendation Since this appears to be a specific behavioral issue not fully covered in the available documentation, I'd suggest reaching out to support directly with your agent ID and call IDs for investigation. Check out these potentially useful sources: - [MCP Documentation]() - [Connect any AI Voice Agent to MCP]() Would you like me to try providing a general answer about LLM tool-call handling patterns?
​ To learn more, see the sources I considered: [[1] MCP](https://docs.retellai.com/build/single-multi-prompt/mcp) [[2] Conversation flow agent stops speaking after calling function](https://discord.com/channels/1186424376300679209/1394277381782638672) [[3] this morning all of my calls started to fail.](https://discord.com/channels/1186424376300679209/1394324743615811665) ​
Feel free to tag @AI Evy with additional questions. I also appreciate feedback, it helps me improve.
PreviousNext
Powered by